/generalizablersc

Cross-dataset Learning for Generalizable Land Use Classification (PyTorch)

Primary LanguagePythonGNU General Public License v3.0GPL-3.0

Cross-dataset Learning for Generalizable Land Use Classification (PyTorch)

Training and testing scripts for few-shot, cross-domain RSC with cross-dataset learning.

Dependencies:

You can use the provided .yml Conda environment file to easily install dependencies into a separate environment:

# From repo root
cd env
conda env create -f rsfinetuning.yml
conda activate rsfinetuning

Setup :

Our code structure is inspired from [Cycle-GAN].

The following datasets can be used for training and testing:

  • AID
  • Brazilian Coffe Scenes
  • PatternNet
  • RESISC45
  • RSICB-256
  • RSSCN7
  • SIRI-WHU
  • UCM
  • WHU-RS19

Dataroots must be hardcoded in the corresponding data/$DATASETNAME_dataset.py file. A script will automatically generate a dataset file in the dataroot folder the first time a dataset is used to avoid systematic parsing.

Training :

Run the following script to train (replace $CHECKPOINTS_DIR with the location where you want to save networks and tensorboard data):

Training on multiple datasets
python3 train.py --batch-size 32 --batch-test 512 --checkpoints-dir $CHECKPOINTS_DIR --dim 512 --gpu-ids 0 --imsize 256 --lr 0.0001 --model deepdesc --net resnet50 --niter 10 --niter-decay 5 --save-epoch-freq 1 --val-freq 1 --test-dataset aid --whiten

Testing :

Test performance can be evaluated while training using the --test-datasets option (multiple datasets possible), or separately using a trained model loaded from a run folder $RUN:

Testing
python3 test.py --model deepdesc --load-from $RUN --net resnet50 --dim 512 --whiten --gpu-ids 0 --test-datasets aid --batch-test 512

Publication :

Have a look at the detailed method and results in our Earthvision 2022 paper (CVPR workshop) that received the Best Paper Award. If you reuse our code or results, please consider citing:

@inproceedings{gominski_cross-dataset_2022,
	title = {Cross-dataset Learning for Generalizable Land Use Scene Classification},
	doi = {10.1109/CVPRW56347.2022.00144},
	booktitle = {2022 {IEEE}/{CVF} Conference on Computer Vision and Pattern Recognition Workshops ({CVPRW})},
	author = {Gominski, Dimitri and Gouet-Brunet, Valérie and Chen, Liming},
	date = {2022-06},
}

Acknowledgments :

Thanks to :